You are on page 1of 35

Review of Economic Studies (1998) 65, 261-294 1998 The Review of Economic Studies Limited

0034-6527/98/00120261$02.00

Matching As An Econometric Evaluation Estimator


JAMES J. HECKMAN University of Chicago HIDEHIKO ICHIMURA University of Pittsburgh and PETRA TODD University of Pennsylvania
First version received October 1994, final version accepted October 1997 (Eds.) This paper develops the method of matching as an econometric evaluation estimator. A rigorous distribution theory for kernel-based matching is presented. The method of matching is extended to more general conditions than the ones assumed in the statistical literature on the topic. We focus on the method of propensity score matching and show that it is not necessarily better, in the sense of reducing the variance of the resulting estimator, to use the propensity score method even if propensity score is known. We extend the statistical literature on the propensity score by considering the case when it is estimated both parametrically and nonparametrically. We examine the benefits of separability and exclusion restrictions in improving the efficiency of the estimator. Our methods also apply to the econometric selection bias estimator.

1. INTRODUCTION Matching is a widely-used method of evaluation. It is based on the intuitively attractive idea of contrasting the outcomes of programme participants (denoted Yi) with the outcomes of "comparable" nonparticipants (denoted yo). Differences in the outcomes between the two groups are attributed to the programme. Let /o and /i denote the set of indices for nonparticipants and participants, respectively. The following framework describes conventional matching methods as well as the smoothed versions of these methods analysed in this paper. To estimate a treatment effect for each treated person ie/i, outcome F,, is compared to an average of the outcomes Yoj for matched persons jelo in the untreated sample. Matches are constructed on the basis of observed characteristics X in R"". Typically, when the observed characteristics of an untreated person are closer to those of the treated person ieli, using a specific distance measure, the untreated person gets a higher weight in constructing the match. The estimated gain for each person i in the treated sample is (1) where lVNo,N,(i,j) is usually a positive valued weight function, defined so that for each ' ^ ^ 1 ' Z/e/ ^No.NXiJ) = ^' and No and A^i are the number of individuals in /o and /i, respectively. The choice of a weighting function reflects the choice of a particular distance measure used in the matching method, and the weights are based on distances in the X
261

262

REVIEW OF ECONOMIC STUDIES

space. For example, for each ieli the nearest-neighhour method selects one individual jelo as the match whose Xj is the "closest" value to Z,, in some metric. The kernel methods developed in this paper construct matches using all individuals in the comparison sample and downweighting "distant" observations. The widely-used evaluation parameter on which we focus in this paper is the mean effect of treatment on the treated for persons with characteristics X.

where Z)= 1 denotes programme participation. Heckman (1997) and Heckman and Smith (1998) discuss conditions under which this parameter answers economically interesting questions. For a particular domain ^ for X, this parameter is estimated by Lei, ^^o-NMiYu-Y^j^,^ W^o.NXi,J)Yoj], (2)

where different values of Wffo.NXi) may be used to select different domains ^ or to account for heteroskedasticity in the treated sample. Different matching methods are based on different weighting functions {w^,,^Xi)} and {fVi^.^NXiJ}} The method of matching is intuitively appealing and is often used hy applied statisticians, but not by economists. This is so for four reasons. First, it is difficult to determine if a particular comparison group is truly comparable to participants (i.e. would have experienced the same outcomes as participants had they participated in the programme). An ideal social experiment creates a valid comparison group. But matching on the measured characteristics available in a typical nonexperimental study is not guaranteed to produce such a comparison group. The published literature presents conditional independence assumptions under which the matched group is comparable, but these are far stronger than the mean-independence conditions typically invoked by economists. Moreover, the assumptions are inconsistent with many economic models of programme participation in which agents select into the programme on the basis of unmeasured components of outcomes unobserved by the econometrician. Even if conditional independence is achieved for one set of X variables, it is not guaranteed to be achieved for other sets of X variables including those that include the original variables as subsets. Second, if a valid comparison group can be found, the distribution theory for the matching estimator remains to be established for continuously distributed match variables Z.' Third, most of the current econometric literature is based on separability between observables and unobservables and on exclusion restrictions that isolate different variables that determine outcomes and programme participation. Separability permits the definition of parameters that do not depend on unobservables. Exclusion restrictions arise naturally in economic models, especially in dynamic models where the date of enrollment into the programme differs from the dates when consequences of the programme are measured. The available literature on matching in statistics does not present a framework that incorporates either type of a priori restriction. Fourth, matching is a data-hungry method. With a large number of conditioning variables, it is easy to have many cells without matches. This makes the method impractical or dependent on the use of arbitrary sorting schemes to select hierarchies of matching variables. (See, e.g. Westat (1980,1982,1984).) In an important paper, Rosenbaum and Rubin (1983) partially solve this problem. They establish that if matching on X is valid, so is matching solely on the probability of selection into the prograrrune PT(D=\\X)1. When the match variables are discrete, the matching estimator for each cell is a mean and consistency and asymptotic normality of the matching estimator are easily established.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

263

P{X). Thus a multidimensional matching problem can be recast as a one-dimensional problem and a practical solution to the curse of dimensionality for matching is possible.^ Several limitations hamper the practical application of their theoretical result. Their theorem assumes that the probability of selection is known and is not estimated. It is also based on strong conditional independence assumptions that are difficult to verify in any application and are unconventional in econometrics. They produce no distribution theory for their estimator. In this paper we first develop an econometric framework for matching that allows us to incorporate additive separability and exclusion restrictions. We then provide a sampling theory for matching from a nonparametric vantage point. Our distribution theory is derived under weaker conditions than the ones currently maintained in the statistical literature on matching. We show that the fundamental identification condition of the matching method for estimating (P-1) is whenever both sides of this expression are well defined. In order for both sides to be well defined simultaneously for all X it is usually assumed that O<P(X)<1 so that Supp iX\D= 1) = Supp {X\D = G). As Heckman, Ichimura, Smith, Todd (1998), Heckman, Ichimura, Smith and Todd (19966) and Heckman, Ichimura and Todd (1997) point out, this condition is not appropriate for important applications of the method. In order to meaningfully implement matching it is necessary to condition on the support common to both participant and comparison groups S, where S=Supp {X\D= 1) n Supp and to estimate the region of common support. Equality of the supports need not hold a priori although most formal discussions of matching assumes that it does. Heckman, Ichimura, Smith and Todd (1998) and Heckman, Ichimura and Todd (1997) report the empirical relevance of this point for evaluating job training programmes. Invoking assumptions that justify the application of nonparametric kernel regression methods to estimate programme outcome equations, maintaining weaker mean independence assumptions compared to the conditional independence assumptions used in the literature, and conditioning on S, we produce an asymptotic distribution theory for matching estimators when regressors are either continuous, discrete or both. This theory is general enough to make the Rosenbaum-Rubin theorem operational in the commonly-encountered case where P^X) is estimated either parametrically or nonparametrically. With a rigorous distribution theory in hand, we address a variety of important questions that arise in applying the method of matching: (1) We ask, if one knew the propensity score, P{X), would one want to use it instead of matching on XI (2) What are the effects on asymptotic bias and variance if we use an estimated value of P? We address this question both for the case of parametric and nonparametric PiX). Finally, we ask (3) what are the benefits, if any, of econometric separability and exclusion restrictions on the bias and variance of matching estimators? The structure of this paper is as follows. Section 2 states the evaluation problem and the parameters identified by the analysis of this paper. Section 3 discusses how matching solves the evaluation problem. We discuss the propensity score methodology of Rosenbaum and Rubin (1983). We emphasize the importance ofthe common support condition
2. They term P(X) the propensity score. For the relationship between propensity score methods and selection models, see Heckman and Robb (1986) or Heckman, Ichimura, Smith and Todd (1998).

264

REVIEW OF ECONOMIC STUDIES

assumed in the literature and develop an approach that does not require it. Section 4 contrasts the assumptions used in matching with the separability assumptions and exclusion restrictions conventionally used in econometrics. A major goal of this paper is to unify the matching literature with the econometrics literature. Section 5 investigates a central issue in the use of propensity scores. Even if the propensity score is known, is it better, in terms of reducing the variance of the resulting matching estimator, to condition on X or P(X)? There is no unambiguous answer to this question. Section 6 presents a basic theorem that provides the distribution theory for kernel matching estimators based on estimated propensity scores. In Section 7, these results are then applied to investigate the three stated questions. Section 8 summarizes the paper. 2. THE EVALUATION PROBLEM AND THE PARAMETERS OF INTEREST Each person can be in one of two possible states, 0 and 1, with associated outcomes (Jo, Yt), corresponding to receiving no treatment or treatment respectively. For example, "treatment" may represent participation in the social programme, such as the job training programme evaluated in our companion paper where we apply the methods developed in this paper. (Heckman, Ichimura and Todd (1997).) Let D= 1 if a person is treated; > = 0 otherwise. The gain from treatment is A= 7 , - yo. We do not know A for anyone because we observe only Y=DYi + {l-D)Yo, i.e. either yo or y,. This fundamental missing data problem cannot be solved at the level of any individual. Therefore, the evaluation problem is typically reformulated at the population level. Focusing on mean impacts for persons with characteristics X, a commonly-used parameter of interest for evaluating the mean impact of participation in social programmes is (P-1). It is the average gross gain from participation in the programme for participants with characteristics X. If the full social cost per participant is subtracted from (P-1) and the no treatment outcome for all persons closely approximates the no programme outcome, then the net gain informs us of whether the programme raises total social output compared to the no programme state for the participants with characteristics X.^ The mean E(Yt\D=l,X) can be identified from data on programme participants. Assumptions must be invoked to identify the counterfactual mean E(Yo\D=l,X), the no-treatment outcome of programme participants. In the absence of data from an ideal social experiment, the outcome of self-selected nonparticipants E(Yo\D = O,X) is often used to approximate E(Yo\D=l,X). The selection bias that arises from making this approximation is

Matching on X, or regression adjustment of yo using X, is based on the assumption that B(X) = O so conditioning on X eliminates the bias. Economists have exploited the idea of conditioning on observables using parametric or nonparametric regression analysis (Barnow, Cain and Goldberger (1980), Barros (1986), Heckman and Robb (1985, 1986)). Statisticians more often use matching methods, pairing treated persons with untreated persons of the same X characteristics (Cochrane and Rubin (1973)). The literature on programme evaluation gives two distinct responses to the problem of estimating (P-1) with continuous conditioning variables. The first borrows from the
3. See Heckman (1997) or Heckman and Smith (1998) for a precise statement of when this parameter answers an mteresting economic evaluation question.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

265

kernel regression literature. It uses a smoothing procedure that borrows strength from adjacent values of a particular value of A'=x and produces uniformly consistent estimators of (P-1) at all points of the support for the distributions oi X given D=\ or D = 0. (See Heckman, Ichimura, Smith and Todd (1996a) or Heckman, Ichimura and Todd (1997).) Parametric assumptions about E(Yo\D=l,X) play the same role as smoothing assumptions, and in addition allow analysts to extrapolate out of the sample for X. Unless the class of functions to which (P-1) may belong is restricted to be smaller than the finiteorder continuously-diiferentiable class of functions, the convergence rate of an estimator of (P-1) is governed by the number of continuous variables included in X (Stone (1982)). The second response to the problem of constructing counterfactuals abandons estimation of (P-1) at any point of A' and instead estimates an average of (P-1) over an interval of X values. Commonly-used intervals include Supp ( X | i ) = l ) , or subintervals of the support corresponding to different groups of interest. The advantage of this approach is that the averaged parameter can be estimated with rate A'^"'''^, where N is sample size, regardless of the number of continuous variables in X when the underlying functions are smooth enough. Averaging the estimators over intervals ofX produces a consistent estimator of M(5) = E ( y , - y o | i ) = l , A ' e 5 ' ) , (P-2)

with a well-defined iV~'^^ distribution theory where S is a subset of Supp {X\D=l). There is considerable interest in estimating impacts for groups so (P-2) is the parameter of interest in conducting an evaluation. In practice both pointwise and setwise parameters may be of interest. Historically, economists have focused on estimating (P-1) and statisticians have focused on estimating (P-2), usually defined over broad intervals of X values, including Supp (Jir|jD= 1). In this paper, we invoke conditions sufficiently strong to consistently estimate both (P-1) and (P-2). 3. HOW MATCHING SOLVES THE EVALUATION PROBLEM Using the notation of Dawid (1979) let

denote the statistical independence of (yo, Yi) and D conditional on X. An equivalent formulation of this condition is This is a non-causality condition that excludes the dependence between potential outcomes and participation that is central to econometric models of self selection. (See Heckman and Honore (1990).) Rosenbaum and Rubin (1983), henceforth denoted RR, establish that, when (A-1) and 0<P{X)<\, (A-2)

are satisfied, (yo, Yi)]LD\P{X), where P(X) = PT (>= 1 IA'). Conditioning on PiX) balances the distribution of yo and Yi with respect to D. The requirement (A-2) guarantees that matches can be made for all values of A T . RR called condition (A-1) an "ignorability" condition for D, and they call (A-1) and (A-2) together a "strong ignorability" condition. When the strong ignorability condition holds, one can generate marginal distributions of the counterfactuals Foiyo\D=l,X) and F,(y,|> = O,Ar),

266

REVIEW OF ECONOMIC STUDIES

but one cannot estimate the joint distribution of (yo, Yi), Fiyo,yi \D, X), without making further assumptions about the structure of outcome and programme participation equations.'* If PiX) = 0 or i'(A') = 1 for some values of Af, then one cannot use matching conditional on those X values to estimate a treatment effect. Persons with such X characteristics either always receive treatment or never receive treatment, so matches from both D = 1 and D = 0 distributions cannot be performed. Ironically, missing data give rise to the problem of causal inference, but missing data, i.e. the unobservables producing variation in D conditional on X, are also required to solve the problem of causal inference. The model predicting programme participation should not be too good so that PiX) = 1 or 0 for any X. Randomness, as embodied in condition (A-2), guarantees that persons with the same characteristics can be observed in both states. This condition says that for any measurable set A, Pr iXeA |) = 1) > 0 if and only if Pr iXeA |) = 0) >0, so the comparison of conditional means is well defined.' A major finding in Heckman, Ichimura, Smith, Todd (1996a, b, 1998) is that in their sample these conditions are not satisfied, so matching is only justified over the subset Supp (Af |Z)= 1) n Supp iX\D = O). Note that under assumption (A-1)

so E(yo|Z>=l,Are5) can be recovered from E(yo|Z) = 0,Ar) by integrating over A T using the distribution ofX given D=l, restricted to S. Note that, in principle, both E(yo|Af, D = 0) and the distribution of A' given D= 1 can be recovered from random samples of participants and nonparticipants. It is important to recognize that unless the expectations are taken on the common support of 5, the second equality does not necessarily follow. While EiYo\D = O,X) is always measurable with respect to the distribution of A T given D = 0, (/i(Ar|Z) = O)), it may not be measurable with respect to the distribution of A T given Z)= 1, (//(Af |) = 1)). Invoking assumption (A-2) or conditioning on the common support S solves the problem because niX\D = O) and //(Af|)=l), restricted to S, are mutually absolutely continuous with respect to each other. In general, assumption (A-2) may not be appropriate in many empirical applications. (See Heckman, Ichimura, and Todd (1997) or Heckman, Ichimura, Smith and Todd (1996a, b, 1998).) The sample counterpart to the population requirement that estimation should be over a common support arises when the set S is not known. In this case, we need to estimate S. Since the estimated set, S, and S inevitably differ, we need to make sure that asymptotically the points at which we evaluate the conditional mean estimator ofEiYo\D = 0,X) are in S. We use the "trimming" method developed in a companion paper (Heckman, Ichimura, Smith, Todd (1996a)) to deal with the problem of determining the points in S. Instead of imposing (A-2), we investigate regions S, where we can reasonably expect to learn about E ( y i - Yo\D=l, S). Conditions (A-1) and (A-2) which are commonly invoked to justify matching, are stronger than what is required to recover E(yi - yo|)= 1, Af) which is the parameter of
4. Heckman, Smith and Clements (1997) and Heckman and Smith (1998) analyse a variety of such assumptions. 5. Thus, the implications of 0<Pr (D= l\X)<\ is that conditional measures of A" given Z) = O and that given D=\ are absolutely continuous with respect to each other. These dominating measure conditions are standard in the matching literature.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

267

interest in this paper. We can get by with a weaker condition since our objective is construction of the counterfactual E(yo|Af, D = 1) (A-3) which implies that Pr(yo</|X>=l,A') = P r ( y o < / | D = 0,Af) for Af65. In this case, the distribution of yo given X for participants can be identified using data only on nonparticipants provided that AfeS. From these distributions, one can recover the required counterfactual mean E(yo|D= 1,A') for AfeS. Note that condition (A-3) does not rule out the dependence of D and Y] or on A = yi - yo given X. * For identification of the mean treatment impact parameter (P-1), an even weaker mean independence condition sufiices E(yo|iD=l,Af) = E(yo|> = 0,Af) for AfeS. (A-1')

Under this assumption, we can identify EiYo\D= 1, X) for XeS, the region of common support.' Mean independence conditions are routinely invoked in the econometrics literature.* Under conditions (A-1) and (A-2), conceptually different parameters such as the mean effect of treatment on the treated, the mean effect of treatment on the untreated, or the mean effect of randomly assigning persons to treatment, all conditional on X, are the same. Under assumptions (A-3) or (A-1'), they are distinct.' Under these weaker conditions, we demonstrate below that it is not necessary to make assumptions about specific functional forms of outcome equations or distributions of unobservables that have made the empirical selection bias literature so controversial. What is controversial about these conditions is the assumption that the conditioning variables available to the analyst are sufficiently rich to justify application of matching. To justify the assumption, analysts implicitly make conjectures about what information goes into the decision sets of agents, and how unobserved (by the econometrician) relevant information is related to observables. (A-1) rules out dependence of D on yo and Yx and so is inconsistent with the Roy model of self selection. See Heckman (1997) or Heckman and Smith (1998) for further discussion.

4. SEPARABILITY AND EXCLUSION RESTRICTIONS In many applications in economics, it is instructive to partition X into two not-necessarily mutually exclusive sets of variables, (T, Z), where the T variables determine outcomes o,
g i ) u

(3a)
(3b)

and the Z variables determine programme participation Pr(Z)=l|Af) = P r ( ) = l | Z ) = /'(Z). (4)

6. By symmetric reasoning, if we postulate the condition YxlD\X and (A-2), then Vr(D=\\Y\,X) = Pr {D= 1 \X), so selection could occur on Jo or A, and we can recover VT(Yi<t\D=Q,X). Since ?r{Ya<t\D = 0,X) can be consistently estimated, we can recover E{Y\- Ya\D=Q,X). 7. We can further identify E ( y , - yol^=0) if we assume E(r,|^=l.^)=E(Ki|/)=O,;f) for A " in S. 8. See, for example, Barnow, Cain and Goldberger (1980) or Heckman and Robb (1985, 1986). 9. See Heckman (1990, 1997) for a discussion of the three parameters. See also Heckman and Smith (1998).

268

REVIEW OF ECONOMIC STUDIES

Thus in a panel data setting y, and yo may be outcomes measured in periods after programme participation decisions are made, so that Z and T may contain distinct variables, although they may have some variables in common. Different variables may determine participation and outcomes, as in the labour supply and wage model of Heckman (1974). Additively-separable models are widely used in econometric research. A major advantage of such models is that any bias arising from observing yo or y, by conditioning on D is confined to the "error term" provided that one also conditions on T, e.g. EiYo\D = \,X)=goiT) + EiUo\D=\,Z) and E(y,|i)= 1,Af)=g,(r)-t-E(:/,|>= 1, Z). Another major advantage of such models is that they permit an operational definition ofthe effect of a change in T holding U constant. Such effects are derived from the ^o and ^i functions. The Rosenbaum-Rubin Theorem (1983) does not inform us about how to exploit additive separability or exclusion restrictions. The evidence reported in Heckman, Ichimura, Smith and Todd (1996a), reveals that the no-training earnings of persons who chose to participate in a training programme, yo, can be represented in the following way where Z and T contain some distinct regressors. This representation reduces the dimension of the matching or nonparametric regression if the dimension of Z is two or larger. Currently-available matching methods do not provide a way to exploit such information about the additive separability of the model or to exploit the information that Z and T do not share all of their elements in common. This paper extends the insights of Rosenbaum and Rubin (1983) to the widely-used model of programme participation and outcomes given by equations (3a), (3b) and (4). Thus, instead of (A-1) or (A-3), we consider the case where UoiLD\X. (A-4a)

Invoking the exclusion restrictions PiX) = PiZ) and using an argument analogous to Rosenbaum and Rubin (1983), we obtain E{)| C/o,/'(Z)} =E{E(Z)| C/o, Af) I = E{P(Z)|t/ so that UolD\PiZ). (A-4b)

Under condition (A-4a) it is not necessarily true that (A-1) or (A-3) are valid but it is obviously true that

In order to identify the mean treatment effect on the treated, it is enough to assume that E(t/o|Z)=l,P(Z)) = E(C/o|i) = 0,P(Z)), (A-4b') instead of (A-4a) or (A-4b). Observe that (A-4a), (A-4b), and (A-4b') do not imply that E(f/o|/'(Z)) = 0 or that E(f/i li'(Z)) = 0. They only imply that the distributions ofthe unobservables are the same in populations of Z)=l and Z) = 0, once one conditions on PiZ). Yo and y, must be adjusted to eliminate the effects of T on outcomes. Only the residuals can be used to exploit the RR conditions. Thus P(Z) is not, in general, a valid instrumental variable.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

269

In order to place these results in the context of classical econometric selection models, consider the following index model setup >=1 = 0 otherwise. If Z and V are independent, then P(Z) = F^(v/(Z)), where Fv( ) is the distribution function of v. In this case identification condition (A-4b') implies

or when F^ is strictly increasing,

LL
(* oo

C >ii(Z)

= \ I Uo
J-oo 'v(Z)

If, in addition, y/iZ) is independent of (f/o, v), and E(f/o) = 0, condition (*) implies
(oo I'V

Uo, v)dvdUo=0,

for any v(Z), which in turn implies EiUo\v = s) = O for any 5 when v^(Z) traces out the entire real line. Hence under these conditions our identification condition implies there is no selection on unobservables as defined by Heckman and Robb (1985, 1986). However, ilfiZ) may not be statistically independent of (f/o, v). Thus under the conditions assumed in the conventional selection model, the identification condition (A-4b') may or may not imply selection on unobservables depending on whether y/iZ) is independent of (Co, v) or not. 5. ESTIMATING THE MEAN EFFECT OF TREATMENT: SHOULD ONE USE THE PROPENSITY SCORE OR NOT? Under (A-l') with S=SuppiX\D = \) and random sampling across individuals, if one knew EiYo\D = 0,X=x), a consistent estimator of (P-2) is

where /i is the set of / indices corresponding to observations for which D, = 1. If we assume E(yol^=l,i'(A')) = E(yo|Z> = 0,P(^))forZ6Supp(/'(^)|Z)=l), (A-l") which is an implication of (A-l), and E(rol^ = O, PiX)=p) is known, the estimator

is consistent for E(A|D= 1). We compare the efficiency of the two estimators, Ap and Ax. We show that neither is necessarily more efficient than the other. Neither estimator is feasible because both assume the conditional mean function and PiX) are known whereas in practice they need

270

REVIEW OF ECONOMIC STUDIES

to be estimated. However, the analysis of this case is of interest because the basic intuition from the simple theorem established below continues to hold when the conditional mean function and P{X) are estimated.
Theorem 1. Assume:

(i) (A-l') and (A-l") hotdfor 5=Supp {X\D= 1);


(ii) {^li,-^1 }/6/, cire independent and identicatty distributed; and

(iii) 0<E(F?) E(y?)<oo. Then Kx and Ap are both consistent estimators of (P-2) with asymptotic distributions that are normat with mean 0 and asymptotic variances Vx and Vp, respectively, where (r, |/)= 1, ^ ) |Z)= 1]-H Var [E(y, - ro|/)= 1, A')|Z)= 1], and |Z)= 1,/>(X))|>= 1]-^ Var [E(r, - ro|/)= 1, P(Z))|Z)= 1]. The theorem directly follows from the central limit theorem for iid sampling with finite second moment and for the sake of brevity its proof is deleted. Observe that E[Var (F, |)= 1,Z)|D= l]^E[Var (r, |/)= 1, i>(Z))|Z)= 1], because X is in general a better predictor than P{X) but Var [E(y, - Y^\D= \,X)\D= 1]^ Var [El(y, - Yo) \D^\, P{X)) \D= 1], because vector X provides a finer conditioning variable than P{X). In general, there are both costs and benefits of conditioning on a random vector X rather than P{X). Using this observation, we can construct examples both where F^g Vp and where K^-^ Vp. Consider first the special case where the treatment effect is constant, that is E{Yi-Yo\D=\,X) is constant. An iterated expectation argument implies that E(y| - yo|Z)= 1, P{X)) is also constant. Thus, the first inequality, F^-g Vp holds in this case. On the other hand, if Yx = m{P{X)) + U for some measurable function m{ ) and / and X are independent, then

which is non-negative because vector X provides a finer conditioning variable than P{X). So in this case Vx^Vp. When the treatment effect is constant, as in the conventional econometric evaluation models, there is only an advantage to conditioning on X rather than on P{X) and there is no cost.'" When the outcome 7, depends on X only through P(X), there is no advantage to conditioning on X over conditioning on P(X). Thus far we have assumed that P(X) is known. In the next section, we investigate the more realistic situation where it is necessary to estimate both P{X) and the conditional
10. Heckman (1992), Heckman and Smith (1998) and Heckman, Smith and Clements (1997) discuss the central role ofthe homogeneous response assumption in conventional econometric models of program evaluation.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

271

means. In this more realistic case, the trade-off between the two terms in Vx and Vp persists." When we need to estimate P{X) or E(yo| Z) = 0, X), the dimensionality of the ^ is a major drawback to the practical application of the matching method or to the use of conventional nonparametric regression. Both are data-hungry statistical procedures. For high dimensional X variables, neither method is feasible in samples of the size typically available to social scientists. Sample sizes per cell become small for matching methods with discrete A"s. Rates of convergence slow down in high-dimensional nonparametric methods. In a parametric regression setting, one may evade this problem by assuming functional forms for E(t/o|A') (see e.g. Barnow, Cain and Goldberger (1980) and the discussion in Heckman and Robb (1985, 1986)), but this approach discards a major advantage of the matching method because it forces the investigator to make arbitrary assumptions about functional forms of estimating equations. Conditioning on the propensity score avoids the dimensionality problem by estimating the mean function conditional on a one-dimensional propensity score P{X). However, in practice one must estimate the propensity score. If it is estimated nonparametrically, we again encounter the curse of dimensionality. The asymptotic distribution theorem below shows that the bias and the asymptotic variance of the estimator of the propensity score affects the asymptotic distribution of the averaged matching estimator more the larger the effect of a change in the propensity score on the conditional means of outcomes. 6. ASYMPTOTIC DISTRIBUTION THEORY FOR KERNEL-BASED MATCHING ESTIMATORS We present an asymptotic theory for our estimator of treatment effect (P-2) using either identifying assumption (A-l') or (A-4b'). The proof justifies the use of estimated P values under general conditions about the distribution of X. We develop a general asymptotic distribution theory for kemel-regression-based and local-polynomial-regression-based matching estimators of (P-2). Let Tand Z be not necessarily mutually exclusive subvectors of ^ , as before. When a function depends on a random variable, we use corresponding lower case letters to denote its argument, for example, g(,t,p) = E{Yo\D=\, T=t,P{Z)=p) and P(z) = Pr (D= 1 |Z=z). Although not explicit in the notation, it is important to remember that g{t, p) refers to the conditional expectation conditional on Z)= 1 as well as T=t and P(Z)=p. We consider estimators of g(r, i'(z)) where P{z) must be estimated. Thus we consider an estimator g^t, P{z)), where P{z) is an estimator of Piz). The general class of estimators of (P-2) that we analyse are of the form ^ NT'L.,[Yu-giTi,PiZd)]IiXieS)
^^'' (6)

where I{A) = l if A holds and =0 otherwise and S is an estimator of S, the region of overlapping support, where S = Supp {X\D=\}r\Supp {X\D = O}. ^ To establish the properties of matching estimators of the form K based on different estimators of P(z) and g{t, Piz)), we use a class of estimators which we call asymptotically
11. If we knew E( y, | Z) = 1, P(X )=p)as well, the estimator would be more efficient than Ap. In practical applications, we do not know either E(Yi\D=\,P{x)=p) E{Yo\D = 0, P(X)=p) so this point is only a theoretical curiosum and is not investigated further. or

272

REVIEW OF ECONOMIC STUDIES

linear estimators with trimming. We analyse their properties by proving a series of lemmas and corollaries leading up to Theorem 2. With regard to the estimators P(z) and git,p), we only assume that they can he written as an average of some function of the data plus residual terms with appropriate convergence properties that are specified below. We start by defining the class of asymptotically linear estimators with trimming. Definition L An estimator Oix) of Oix) is an asymptotically linear estimator with trimming IixeS) if and only if there is a function y/", defined over some subset of a finite-dimensional Euclidean space, and stochastic terms Six) and Rix) such that for sample size n:

(i) [9ix)-9ix)]IixeS) = n-' Yl^^ ^|/iXi, Yr,x) + Six (ii) E{yfiXl,Yr,X)\X=x} = O, (iii) p\im^^n-'/^X"=, KX,) = b<oo;
An estimator ^ of )3 is called asymptotically linear if

holds.'^ Definition 1 is analogous to the conventional definition, hut extends it infiveways to accommodate nonparametric estimators. First, since the parameter Oix) that we estimate is a function evaluated at a point, we need a notation to indicate the point x at which we estimate it. Conditions (i)-(iv) are expressed in terms of functions of x. Second, for nonparametric estimation, asymptotic linearity only holds over the support of Xthe region where the density is hounded away from zero. To define the appropriate conditions for this restricted region^ we introduce a trimming function IixeS) that selects ohservations only if they lie in S and discards them otherwise. Third, nonparametric estimators depend on smoothing parameters and usually have bias functions that converge to zero for particular sequences of smoothing parameters. We introduce a subscript n to the ^-function and consider it to he an element of a class of functions ^n, instead of a fixed function, in order to accommodate smoothing parameters. For example, in the context of kernel estimators, if we consider a smoothing parameter of the form ai^x) hn, different choices of / generate an entire class of functions H'n indexed hy a function c( ) for any given kernel." We refer to the function Vn as a score function. The stochastic tenn^^(j:) is the hias term arising from estimation. For parametric cases, it often happens that Z>(jc) = O. Fourth, we change the notion of the residual term heing "small" from Op(/i~'^^) to the weaker condition (iv). We will demonstrate that this weaker condition is satisfied hy some nonparametric estimators when the stronger condition 0;,(n~'^^) is not. Condition (iii) is required to restrict the behaviour of the hias term. The hias term has to be reduced to a rate o( '^^) in order to properly centre expression (i) asymptotically. For the case ofrf-dimensionalnonparametric model with/?-times continuously differentiable functions. Stone (1982) proves that the optimal uniform rate of convergence of the nonparametric regression function with respect to mean square error is (/log n)'''^^^'''"^. His result implies that some undersmoothing, compared to this optimal rate, is required to achieve the desired rate of convergence in the hias term alone. Note that the higher the dimension of the estimand, the more adjustment in smoothing parameters to reduce bias is required.
12. See Newey and McFadden (1994) p. 2142. 13. See Ichimura (1996).

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

273

This is the price that one must pay to safeguard against possible misspecifications of g(?, p) or Piz). It is straightforward to show that parametric estimators of a regression function are asymptotically linear under some mild regularity conditions. In the Appendix we establish that the local polynomial regression estimator of a regression function is also asymptotically linear. A typical estimator of a parametric regression function mix; P) takes the form w(x; P), where m is a known function and P is an asymptotically linear estimator, with ^ n''^^^). In this case, by a Taylor expansion.

, P)-mix, p)] = n-'/' Y.%, [dmix, p)/8pWiX,, 7,) + [dmix, P)/dp-8mix, '

where P lies on a line segment between P and p. When E{ii/iXi, Yi)}=0 and E{\i/iXi, Yi){i/iXi, Y,)'}<oo, under iid sampling, for example, n~'^^EJ=i ViXi, Yi) =

Opil) and plim^o,p = P so that p\imr,^^\dmix, P)/dp-dmix, P)/8p\=0j,il) if


dmix, P)/dp is Holder continuous at p.^* Under these regularity conditions ^[mix, p)-mix, )3)] = -/'X-=, [^'(^' P)/8PMX,, y,) + o^(l).

The bias term of the parametric estimator mix, P) is Six) = 0, under the conditions we have specified. The residual term satisfies the stronger condition that is maintained in the traditional definition of asymptotic linearity.

(a) Asymptotic linearity of the kernel regression estimator We now establish that the more general kernel regression estimator for nonparametric functions is also asymptotically linear. Corollary 1 stated below is a consequence of a more general theorem proved in the Appendix for local polynomial regression models used in Heckman, Ichimura, Smith and Todd (1998) and Heckman, Ichimura and Todd (1997). We present a specialized result here to simplify notation and focus on main ideas. To establish this result we first need to invoke the following assumptions. Assumption 1. Sampling of {Xi, 7,} is i.i.d., X, takes values in R!' and F, in R, and Var(y,)<oo. When a function is />-times continuously differentiable and its p-th derivative satisfies Holder's condition, we call the function/7-smooth. Let mix) = E{Yi\Xi=x}. Assumption 2. mix) is ^-smooth, where p>d.
14. A function is Holder continuous at A'=xo with constant 0 < a g 1 if | q^x, 0) - (p{xo, fl)| g C \\x-Xo\\ for some C > 0 for all x and 0 in the domain of the function cp(-,-). Usually Holder continuity is defined for a function with no second argument B. We assume that usual Holder continuity holds uniformly over 0 whenever there is an additional argument.

274

REVIEW OF ECONOMIC STUDIES We also allow for stochastic bandwidths:

Assumption 3. Bandwidth sequence a, satisfies p\im^^a/h = ao>0 for some deterministic sequence {h} that satisfies nh^/logn-yao and nhf-*c<ao for some c^O. This assumption implies 2^>rf but a stronger condition is already imposed in Assumption 2.'^ Assumption 4. Kernel function ^( ) is symmetric, supported on a compact set, and is Lipschitz continuous. The assumption of compact support can be replaced by a stronger assumption on the distribution of Xi so that all relevant moments exist. Since we can always choose Ki ), but we are usually not free to pick the distribution of Z,, we invoke compactness. In this paper we consider trimming functions based on S and S that have the following structure. hQifxix) be the Lebesgue density oiX,, S= {xeR'';fxix)^qo}, and S= {xeR''; fxix)^qo}, where supxes\fxix)-fxix)\ converges almost surely to zero, and/;i-(jc) ispsmooth. We also require that fxix) has a continuous Lebesgue density j|^ in a neighbourhood of qo vnthffiqo) > 0. We refer to these sets S and S as p-nice on S. The smoothness of fxix) simplifies the analysis and hence helps to establish the equicontinuity results we utilize. Assumption 5. Trimming is ^-nice on S. In order to control the bias of the kernel regression estimator, we need to make additional assumptions. Certain moments of the kernel function need to be 0, the underlying Lebesgue density of Xt,fxix), needs to be smooth, and the point at which the function is estimated needs to be an interior point ofthe support of Z,. It is demonstrated in the Appendix that these assumptions are not necessary for ^-th order local polynomial regression estimator. Assumption 6. Kernel function Ki ) has moments of order 1 through p - 1 that are equal to zero. Assumption!. />(:>(:) is ^-smooth. Assumption 8. A point at which m(-) is being estimated is an interior point of the support of A',. The following characterization of the bias is a consequence of Theorem 3 that is proved in the Appendix.
CoroUary 1. Under Assumptions 1-8, if Kiu^,..., Ud)=kiui) kiuj) where ki ) is a one dimensional kernel, the kernel regression estimator rhoix) of mix) is asymptotically
15. Assumption 3 implies /> - 0 and log n A ^^ "'' - 0. These two imply 2p > d. Also notice that the assumption implies K-^O.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

275

linear with trimming., where, writing e,= Yi-B{Yi\Xi}, and Wn(.Xi, Yi x) = (naohi)-'EiKi{Xi-x)/iaohr.))r{xeS)/fx(x),

Six) =

Our use of an independent product form for the kernel function simplifies the expression for the bias function. For a more general expression without this assumption see the Appendix. Corollary 1 differs from previous analyses in the way we characterize the residual term. (b) Extensions to the case of local polynomial regression In the Appendix, we consider the more general case in which the local polynomial regression estimator for g(t, p) is asymptotically linear with trimming with a uniformly cqiisistent derivative. The latter property is useful because, as the next lemma shows, if both P{z) and g{t, p) are asymptotically linear, and if dg^t, p)/dp is uniformly consistent, then g^t, P(z)) is also asymptotically linear under some additional conditions. We also verify in the Appendix that these additional conditions are satisfied for the local polynomial regression estimators. Let P,(z) be a function that is defined by a Taylor's expansion pf i(r, P(z)) in the neighbourhood of P(z), i.e. g{t, P(,z))=g(,t, P(z)) + dg(t, P,(z))/dp [P(z)-P{z)l Lemma 1. Suppose that: (i) Both P(z) andg(t,p) are asymptotically linear with trimming where

(ii) dg(t,p)/dp and P(z) are uniformly consistent and converge to dg{t,p)/dp and P(z), respectively and dg(t,p)/dp is continuous; (iii) plim-.oo "'/' i ; - , S,(T,, P(Z,)) = b, and plim^-'/' i;_, 8g(Ti, PiZ,))/8p b,iT,, P(Z,)) = b,^; ^ (iv) plim^.oo "'^' Z;-, lSg(T>, PT.iZi))/Sp-8g(Ti, P(Z,))/8p] (V) plim^^^ n-'^' i ; ^ , i ; 8 d i j P{Zd)/ then g(t, P{z)) is also asymptotically linear where [g(t, Piz))-g{t, PizW(xeS) = n-' S;=, IVnAYj, Tj, P{Zj); t, P(z)) + dgit, P(z))/dp yifnp^Dj, Zj; z)] -I-b{x) + R{x),

andli

~'^^ Z" MX)

276

REVIEW OF ECONOMIC STUDIES

An important property of this expression which we exploit below is that the effect of the kernel function v^n^CA, 'Zj\'^) always enters multiplicatively with dg{t, P{z))/dp. Thus both the bias and variance of P{z) depend on the slope to g with respect to p. Condition (ii) excludes nearest-neighbour type matching estimators with a fixed number of neighbours. With conditions (ii)-(v), the proof of this lemma is just an application of Slutsky's theorem and hence the proof is omitted. In order to apply the theorem, however, we need to verify the conditions. We sketch the main arguments for the case of parametric estimator P(z) of P{z) here and present proofs and discussion of the nonparametric case in the Appendix. ^ Under the regularity conditions just presented, the bias function for a parametric P{z) is zero. Hence condition (iii) holds '\fg(,t,p) is asymptotically linear and its derivative is uniformly consistent for the true derivative. Condition (iv) also holds since \Rp{Z,) | = Op {n '/^) and the derivative oig{t, p) is uniformly consistent. Condition (v) can be verified by exploiting the particular form of score function obtained earlier. Observing that
W{D Z; Z) v'(Z) WpiiDj, Zf), we obtain

, P{Zi))/dp] g(Ti, P{Zi))/dp]xifp,{Zi)n-''^ Y.%, VpiiDj, Zj), so condition (v) follows from an application of the central limit theorem and the uniform consistency of the derivative of g(t,p). For the case of nonparametric estimators, Vnp does not factor and the double summation does not factor as it does in the case of parametric estimation. For this more general case, we apply the equicontinuity results obtained by Ichimura (1995) for general Ustatistics to verify the condition. We verify all the conditions for the local polynomial regression estimators in the Appendix. Since the derivative of g{t,p) needs to be defined we assume Assumption 9. K() is 1-smooth. Lemma 1 implies that the asymptotic distribution theory of A can be obtained for those estimators based on asymptotically linear estimators with trimming for the general nonparametric (in P and g) case. Once this result is established, it can be used with lemma 1 to analyze the properties of two stage estimators of the form g(t, P(z)). (c) Theorem 2: The asymptotic distribution of the matching estimator under general conditions Theorem 2 enables us to produce the asymptotic distribution theory of a variety of estimators A of the treatment effect under different identifying assumptions. It also produces the asymptotic distribution theory for matching estimators based on different estimators of g(t,p) and P(z). In Sections 3 and 4, we presented various matching estimators for the mean effect of treatment on the treated invoking different identifying assumptions. An altemative to matching is the conventional index-sufficient selection estimators that can be used to construct estimators of E (yo|/)= 1, A^), as described in our companion paper Heckman, Ichimura and Todd (1997) and in Heckman, Ichimura, Smith and Todd (1996a, 1998). Our analysis is sufficiently general to cover the distribution theory for that case as well provided an exclusion restriction or a distributional assumption is invoked.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

277

Denote the conditional expectation or variance given that A' is in S by Es ( ) or Var5( ), respectively. Let the number of observations in sets h and h be No and N\, respectively, where N=No + N^ and that 0<limAr_oo Ni/No=0<co. Theorem 2. Under the following conditions: (i) {yo,, Xi },e/o and {Yu, A', },e/, are independent and within each group they are i.i.d. and Yoi for ielo and Yufor ielx each has a finite second moment; (ii) The estimator g{x) of g{x) = ^{Yo\Di=\,Xi=x] is asymptotically linear with trimming., where Mix)-g{x)]I{xeS} = No' !,,, ^foN^sX J'o,, Xi ; x)

and the score functions y/dNoN,(Yd,X; x) for d=0 and 1, the bias term bgix), and the trimming function satisfy: (ii-a) E {Vrf^vo^v,(y</MXi;X)\Di = d,X,D=\)}=Ofor d=0 and 1, and Var {ii/ciNoNXYdi,Xi ,X)} = oiN) for each ielou/, ; (ii-b) pUm^v.^co A^r'^' I , , , , kXi) = b; (ii-c) plim^,^<:o Var {E[y/oNoNXYo>,Xi;X)\ Yoi, Di = O,Xi,D=l]|/>= 1} = Ko<oo plim^v.^oo Var {E[y/,!,,^,iYu, Xi; X) \ Yu, Di= 1, Xi, D= 1] |Z)= 1} = K, < co, and lim^v.^co E{[Yu-giXi)]IiXieS) lXD (ii-d) S and S are p-nice on S, where p>d, where d is the number of regressors in X and fix) is a kernel density estimator that uses a kernel function that satisfies Assumption 6. Then under (A-T) the asymptotic distribution of

is normal with mean ib/Pr iXeS\D=l))

and asymptotic variance

YAT, PiZ), D= + PriXeS\D=l)-^{V,+2Proof. See the Appendix. Theorem 2 shows that the asymptotic variance consists of five components. The first two terms are the same as those previously presented in Theorem 1. The latter three terms are the contributions to variance that arise from estimating gix) = E{Yi\Di=l,Xi=x}. The third and the fourth terms arise from using observations for which D = 1 to estimate gix). If we use just observations for which D = 0 to estimate gix), as in the case of the simple matching estimator, then these two terms do not appear and we only acquire the

278

REVIEW OF ECONOMIC STUDIES

fifth term.'* We consider the more general case with all five terms. If A ^ o is much larger than AT,, then the sampling variation contribution of the Z) = 0 observations is small as 6 is small. Condition (i) covers both random and choice-based sampling and enables us to avoid degeneracies and to apply a central limit theorem. Condition (ii) elaborates the asymptotic linearity condition for the estimator of gix). We assume/7-nice trimming. The additional condition on the trimming function is required to reduce the bias that arises in estimating the support. Note that there is no need for gix) to be smooth. A smoothness condition on gix) is used solely to establish asymptotic linearity of the estimator of gix). Also note that the sampling theory above is obtained under mean independence Strong ignorability conditions given by (A-l), (A-2) or (A-3), while conventional in the matching literature, are not needed but they obviously imply these equalities. Theorem 2 can be combined with the earlier results to obtain an asymptotic distribution theory for estimators that use git, Piz)). One only needs to replace functions iYoi>Xi; x) and iifiN^XYuyXr, x) and the bias term by those obtained in Lemma 1. 7. ANSWERS TO THE THREE QUESTIONS OF SECTION 1 AND MORE GENERAL QUESTIONS CONCERNING THE VALUE OF A PRIORI INFORMATION Armed with these results, we now investigate the three questions posed in the Section 1. il) Is it better to match on PiX) or X if you know Matching on X, Ax involves ^-dimensional nonparametric regression function estimation whereas matching on PiX), Kp only involves one dimensional nonparametric regression function estimation. Thus from the perspective of bias, matching on PiX) is better in the sense that it allows ^A^-consistent estimation of (P-2) for a wider class of models than is possible if matching is performed directly on X. This is because estimation of higherdimensional functions requires that the underlying functions be smoother for bias terms to converge to zero. If we specify parametric regression models, the distinction does not arise if the model is correctly specified. When we restrict consideration to models that permit >/F-consistent estimation either by matching on PiX) or on X, the asymptotic variance of Kp is not necessarily smaller than that of Kx. To see this, consider the case where we use a kernel regression for the D = 0 observations i.e. those with ielo. In this case the score function
w (Y ...,^_ WONoN, ( -fOl, ^i , X)

where ,= yo,-E{ I'o/IA',, A = 0} and we write/A-(x|Z) = O) for the Lebesgue density of X, given Z), = 0. (We use analogous expressions to denote various Lebesgue densities.) Clearly V] and Covi are zero in this case. Using the score function we can calculate K o when we match on X. Denoting this variance by Vox,
8(x).
16. An earlier version ofthe paper assumed that only observations for which ) = 0 are used to estimate

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

279

M'

Now observe that conditioning on Xi and Fo,, e, is given, so that we may write the last expression as

I r^((^r:^)^^^Mf^
r UjxiX\D = O)JKiu)du^
Now

1 ij
J' 1
j

r/.((X-^)/a.)/(X.5)
d'

can be written in the following way, making the change of variable iX/aNoW\D = 0)

Taking limits as No-^oa, and using assumptions 3, 4 and 7, so we can take limits inside the integral

since OATO^O and ^Kiw)dw/lKiu)du= 1. Thus, since we sample the Xi for which A = 0,

./A-(.^/IM "J

Hence the asymptotic variance of AA- is, writing A=Pr {A'e5|Z) = 0}/Pr (.YeS'|)= 1), Pr iXeS\D = l)-'{YaTs [EsiY,- Yo\X, D= 1)\D= l] + Es[Vars(F, \X, Z)=

Similarly for Ap, VOP is

The first two terms for both variance expressions are the same as those that appear in Vx and Vp in Theorem 1. To see that one variance is not necessarily smaller than the other, consider the case where fxiX\D=l)=fxiX\D = O) and 9 = 1. Clearly in this case fpiPiX)\D= l)=fpiPiX)\D = O). Propensity score matching has smaller variance if and

280 only if

REVIEW OF ECONOMIC STUDIES

Es{EsiY,\PiX),D=l)EsiYo\PiX),D=l)\D=l} >Es{EsiYt\X,D=l)EsiYo\X,D=l)\D=\}. Since the inequality does not necessarily hold, the propensity score matching estimator in itself does not necessarily improve upon the variance ofthe regular matching estimator."'* (2) What are the effects on asymptotic bias and variance if we use an estimated value of F>. When Pix) is estimated nonparametrically, the smaller bias that arises from matching on the propensity score no longer holds true if estimation of Pix) is a ^/-dimensional nonparmetric estimation problem where rf> 1. In addition, estimation of Pix) increases the asymptotic variance. Lemma 1. informs us that the score, when we use estimated Piz) but no other conditioning variables, is

for ieIj,d=O,\, where \i/dNoN,g are the scores for estimating gip) and IJ/NP is the score for estimating Piz). By assumption (ii-a) they are not correlated with dgiPiz))/8pWspiDj, Zj; z), and hence the variance ofthe sum ofthe scores is the sum ofthe variances of each score. So the variance increases by the variance contribution of the score SgiPiz))/dp y/NpiDj, Zjiz) when we use estimated, rather than known, Piz). Even with the additional term, however, matching on X does not necessarily dominate matching on PiX) because the additional term may be arbitrarily close to zero when dgiPiz))/8p is close to zero. (3) What are the benefits, if any, of econometric separability and exclusion restrictions on the bias and variance of matching estimators? We first consider exclusion restrictions in the estimation of Pix). Again we derive the asymptotic variance formulae explicitly using a kernel regression estimator. Using Corollary 1, the score function for estimating Pix) is a%fxix)\Kiu)du '

uj=Dj-E{Dj\Xj}. Hence the variance contribution of estimation of Piz) without imposing exclusion restrictions is V2x= lim y&r{E[dgiPiZ))/dp
N*co

Kiu)du]\Dj,Xj,D=l]} iDj\Xj)[dgiPiZj))/dp]' -fliXjlD^ l)/f%iXj)][PT {XjeS}]-\ Analogously, we define the variance contribution of estimating P(z) imposing exclusion restrictions by Viz. Observe that when Z is a subset of the variables in X, and when there
17. FoTexamp\eEs(Y,\X,D\) Es{Yo\X,D=\)oTEs(Y,\X,D=\) = -Es(Yo\X,D=\)ainho\d. 18. In a later analysis, Hahn (1996) considers a special case of the models considered in this paper and shows in a model with no exclusion restrictions that when P is not known, the estimated propensity score estimator is efficient and that knowledge of P improves efficiency.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

281

are exclusion restrictions so PiX) = PiZ) then one can show that V2zS Vix- Thus, exclusion restrictions in estimating PiX) reduce the asymptotic variance of the matching estimatoran intuitively obvious result. To show this first note that in this case Var (D|^) = Var (/)|Z). Thus

iD\Z)[dgiPiZ))/8pf

= 0.

Since the other variance terms are the same, imposing the exclusion restriction helps to reduce the asymptotic variance by reducing the estimation error due to estimating the propensity score. The same is true when we estimate the propensity score by a parametric method. It is straightforward to show that, holding all other things constant, the lower the dimension of Z, the less the variance in the matching estimator. Exclusion restrictions in T also reduce the asymptotic variance of the matching estimator. By the same argument, it follows that E{[f%iX\D=l)/fJciX\D = O)] - 1 |D = 0} ^ 0 . This implies that under homoskedasticity for Fo, the case/(A'|Z)= l)=/(X|> = 0) yields the smallest variance. We next examine the consequences of imposing an additive separability restriction on the asymptotic distribution. We find that imposing additive separability does not necessarily lead to a gain in efficiency. This is so, even when the additively separable variables are independent. We describe this using the estimators studied by Tjostheim and Auestad (1994) and Linton and Nielsen (1995)." They consider estimation of fi(^i),g2(A'2) in

where x = (xi, X2). There are no overlapping variables among Xi and X2. In our context, E( Y\X = x)=git) + KiPiz)) and EiY\X=x) is the parameter of interest. In order to focus on the effect of imposing additive separability, we assume /'(z) to be known so that we write P for/'(Z). Their estimation method first estimates E{Y\T=t,P=p}=git) +Kip) non-parametrically, say by fi {y| T= t, P=p}, and then integrates { y| 7 = /, P=p} over p using an estimated marginal distribution of P. Denote the estimator by git). Then under additive separability, git) consistently estimates git)+ E{KiP)}. Analogously one can integrate fi { y| T= t, P=p] git) over t using an estimated marginal distribution of T to obtain a consistent estimator of Kip)-E{KiP)). We add the estimators to obtain the estimator of EiY\X=x) that imposes additive separability. The contribution of estimation of the regression function to asymptotic variance when T and P are independent and additive
19. The derivative of E{ Y\T=t, P=p) with respect to p only depends on p if it is additively separable. Fan, Hardle and Mammen (1996) exploits this property in their estimation. Using this estimator does not lead to an improvement in efficiency, either.

282

REVIEW OF ECONOMIC STUDIES

separability is imposed, is Pr iXeS\D = l)~^ times YolT, P, D =

iP\D = O) fiT\D = O) times

When additive separability is not used, it is Pr iXeS\D=l)~^

Note that the first expression is not necessarily smaller, since fiP\D= 1) /(r|Z>= 1) can be small without both/(P|Z)= 1) and fiT\D=l) being simultaneously small.^' Imposing additive separability per se does not necessarily improve efficiency. This is in contrast to the case of exclusion restrictions where imposing them always improved efficiency. Whether there exists a method that improves efficiency by exploiting additive separability is not known to us. Note that when/(P|Z)= 1)=/(P|D = O) andfiT\D= l ) = / ( r | ) = O) both hold, the variance for the additively separable case and for the general case coincide. Under homoskedasticity of ^0, the most efficient case arises when the distributions of (r, PiZ)) given D= 1 and (T, PiZ)) given D = 0 coincide. In the additively separable case, only the marginal distributions of PiZ) and T respectively have to coincide, but the basic result is the Note that nearest neighbour matching "automatically" imposes the restriction of balancing the distributions of the data whereas kernel matching does not. While our theorem does not justify the method of nearest neighbour matching, within a kernel matching framework we may be able to reweight the kernel to enforce the restrictions that the two distributions be the same. That is an open question which we will answer in our future research. Note that we clearly need to reweight so that the homoskedasticity condition holds. 8. SUMMARY AND CONCLUSION This paper examines matching as an econometric method for evaluating social programmes. Matching is based on the assumption that conditioning on observables eliminates selective differences between programme participants and nonparticipants that are not correctly attributed to the programme being evaluated. We present a framework to justify matching methods that allows analysts to exploit exclusion restrictions and assumptions about additive separability. We then develop a sampling theory for kernel-based matching methods that allows the matching variables to be generated regressors produced from either parametric or nonparametric estimation methods. We show that the matching method based on the propensity score does not
20. The derivation is straightforward but tedious. Use the asymptotic linear representation of the kernel regression estimator and then obtain the asymptotic linear expression using it 21. Let a(P)=f(P\D=l)/f{P\D = O) and b(T)=/{T\D=\)/f{T\D = O) and define an interval H(T) = [l\-b(T)]/[l+b(T)], 1] when b(T)<l. If whenever b(T)>\, a(P)>l and whenever b(T)<\,a{P)eHiT) holds, imposing additive separability improves efficiency. On the other hand, if whenever b(T)>l,a(P)<l and whenever b(T)<\, a(P) lies outside the interval H(T), then imposing additive separability using the available methods worsens efficiency even if the true model is additive. 22. The expression above implies that the same can be said for the estimator that is constructed without imposing additive separability. However that result is an artifact of assuming independence of P(Z) and T.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

283

necessarily reduce the asymptotic bias or the variance of estimators of M(S) compared to traditional matching methods. The advantage of using the propensity score is simplicity in estimation. When we use the method of matching based on propensity scores, we can estimate treatment effects in two stages. First we build a model that describes the programme participation decision. Then we construct a model that describes outcomes. In this regard, matching mimics features of the conventional econometric approach to selection bias. (Heckman and Robb (1986) or Heckman, Ichimura, Smith and Todd (1998).) A useful extension of our analysis would consider the small sample properties of alternative estimators. In samples of the usual size in economics, cells will be small if matching is made on a high-dimensional X. This problem is less likely to arise when matching is on a single variable like P. This small sample virtue of propensity score matching is not captured by our large sample theory. Intuitively, it appears that the less data hungry propensity score method would be more efficient than a high dimensional matching method. Our sampling theory demonstrates the value of having the conditional distribution of the regressors the same for > = 0 and D=\. This point is to be distinguished from the requirement of a common support that is needed to justify the matching estimator. Whether a weighting scheme can be developed to improve the asymptotic variance remains to be investigated. APPENDIX
In this Appendix we prove Corollary 1 by proving a more general result, Theorem 3 stated below, verify the conditions of Lemma 13, for the case of a local polynomial regression estimator, and prove Theorem 2. We first establish the property that local polynomial regression estimators are asymptotically linear with trimming.

A. 1. Theorem 3 Theorem 3 will show that local polynomial regression estimators are asymptotically linear with trimming. Lemma 1 follows as a corollary. The local polynomial regression estimator of a function and its derivatives is based on an idea of approximating the function at a point by a Taylor's series expansion and then estimating the coefficients using data in a neighbourhood ofthe point. In order to present the results, therefore, wefirstdevelop a compact notation to write a multivariate Taylor series expansion. Let x = ( x i , . . . , x^) and 9 = ( 9 , , . . . , qd)eR'' where qj (j= 1 </) are nonnegative integers. Also let y = x?'.. .xy/(q^ ... qJ). Note that we include iq,<.... qjl) in the definition. This enables us to study the derivative of x' without introducing new notation; for example, 5x'/5xi =x* where q = (q,-\ qj), if gi S 1 and 0 otherwise. When the sum of the elements of 9 is i, x* corresponds to a Taylor series polynomial associated with the term d'm(x)/(dx^'... dx^J). In order to consider all polynomials that correspond to i-th order derivatives we next define a vector whose elements are themselves distinct vectors of nonnegative integers whose elements sum to .s. We denote this row vector by Q(s) = ({q],... ,qd))<,,+ +<,^-s', that is Q(s) is a row vector of length (s + d- \y./[s\(d-1)!] whose typical element is a row vector ( 9 , , . . . , qj), which has arguments that sum to .s. For concreteness we assume that {(q,,..., q^)} are ordered according to the magnitude of ^f.., O''"''?; from largest to smallest. We define a row vector x^''' = (x'" "'),, + ...+,^.,. This row vector corresponds to the polynomial terms of degree j. Let x^' = (x'^''^),e{o.i...j,) This row vector represents the whole polynomial up to degree p from lowest to the highest. Also let m'''(x) for s^\ denote a row vector whose typical element is 5'm(x)/(5x1f'... SxJ*) where q, + - +qj=s and the elements are ordered in the same way {{q,,... ,qj)} are ordered. We also write m''(x) = m(x). Let P*(xo) = (m""(xo), . . , m^\xo)y. In this notation, Taylor's expansion of m(x) at Xo to orderp without the remainder term can now be written as (x-xo)^'^*(xo). We now define the local polynomial regression estimator with a global smoothing parameter aA,, where o e [ a o - 5 , a o + 5 ] f o r s o m e a o > 0 a n d a o > 5 > 0 . Wedenoteja^ = [ a o - 5 , a o + 5 ] . Let K|,{s) = {ah)'''K(s/(ah)) and KM(S) = (aM'''. K(s/(aM), and let ^ = ( ) 8 i , . . . , P,)', where p', is conformable with m""(xo), for r = 0 p.

284

REVIEW OF ECONOMIC STUDIES


K.iX^-xo)), and

Also let y = ( y , , . . . , )', W(A:o) = diag (K,(X,-xo),...,

Then the local polynomial regression estimator is defined as the solution to

or, more compactly P,(xo) = arg min (Y-X^(xo)p)' fV(xo)( Y-X,(xo)P). Clearly the estimator equals [X'p(xaW(xo)Xp(xo)Y'X'p{xa)W(xo)Y when the inverse exists. When /) = 0, the estimator is the kernel regression estimator and when/)= 1 the estimator is the local linear regression estimator." Let //=diag (\,(aKyUj,..., (a/i,)"''i(;,+rf-i),/[;,t(rf-i,,,), where i, denotes a row vector of size s with 1 in all arguments. Then P,(x<,) = H[M,{xor'\n''HX',{x)W{xo)Y, where M,Axa) = n-'WX;(xo)W(x<,)X,(xa)H. Note that by Taylor's expansion of order p^p at Xo,m(X,) = (X,-Xn)'P*{xo) + rf(X,,Xo), where rp(.Xi, Xa) = (X,-xo)^''\m'^\x,)-m^\xo)] and Xt lies on a line between X, and Xa. Write
W = (TO(X,), . . . , m(X)), r^(xo) = (r^(X,, ATo),..., r^(X, Xo))', and e = (e^,..., ej'.

Let M,(xo) be the square matrix of size j ; ^ . ^ [{q + d- 1 )\/q\ (d- 1)!] denoting the expectation of //, (xo), where the j-th row, /-th column "block" of A/,(;co) matrix is, for Og.s, r g p .

Let lim,^ M,(xo) = M,/(Jro)''*^ Note that M^ only depends on K() when Xo is an interior point of the support of A-. Also write l = I{X,eS} and /, = /{X,65}, /o = /{A:oeS} and 1^ = 1 {xoeS}. We prove the following theorem.

Theorem 3. Suppose Assumptions 1-4 hold. If M^ is non-singular, then the local polynomial regression estimator of order p, m^ix), satisfies

where 6(xo) = o(/i^),/i-'^';_, R(X,) = o^O), and ; Xo) = (l, 0 0)

Furthermore, suppose that Assumptions 5-7 hold. Then the local polynomial regression estimator of order O^p<p ihpix), satisfies

where

Fan (1993), Ruppert and Wand (1994), and Masry (1995) prove pointwise or uniform convergence property^of the estimator. As for any other nonparametric estimator the convergence rate in this sense is slower than n -rate. We prove that the averaged pointwise residuals converges to zero faster than n"'^^-rate. We only specify that M, is nonsingular because one can find different conditions on Ar( ) to guarantee it. For example, assuming that fC(u,,... ,uj)=k(u,)... k{uj), if J s'^k(s)ds > 0, then M, is nonsingular. 23. Ruppert and Wand (1994) develop multivariate version of the local linear estimator. Masry (1995) develops multivariate general order local polynomial regression.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

285

To prove the theorem, note first that Y=m+B=Xp{xa)P*(xo) + rp{X,,Xa) +e. We wish to consider the situation where the order of polynomial terms included,/?, is less than the underlying smoothness of the regression function, p. For this purpose let Xp(x,,) = [Xp(xQ),Xp(xo)'\ and partition p* conformably: P*(xo) = [^; (xo)', Pt (xo)T. thus note that (A-3) ) h (B-3) (C-3) Note that if we use a p-th order polynomial when m(x) is ;7-th order continuously differentiable, that is p=p, the second term (B-3) is zero. Denote the first element of P{xo) by m^ixo). We establish asymptotic linearity of thpixo) and uniform consistency of its derivative. Lemma 2 shows that the first term (A-3) determines the asymptotic distribution. Lemma 7 shows that the second term (B-3) determines the bias term, and the right hand side of 8 shows that the third term (C-3) is of sufficiently small order to be negligible. Together, these lemmas prove the theorem. Lemma 2 (Term (A-3)). Under the assumptions of Theorem 3, the right-hand side of Io+R^Xa),

(A-3) = e, [M^(xo)Y'"~'H'X'p(X(,)W{xo)e where e, = (1, 0 , . . . , 0) and n"'^' E"-, RAX,) = Op(\).

Proof We first define neighbourhoods of functions e, [Mp{x)Y\fx(x), denote them by Tn, Jtf, J, and si, respectively, where

l(xeS),

and point Oo. We

for some small e , > 0 , j r = {/(x);supsl/(Jc)-/A-(x)l / } for some small e/>0,

J={l(x^S),S=

{x; f{x)^qo} forsome/(A:)e.?r(jc)},

and J3^ = [ff0-5o, ao + 5a], where O<Sa<ao.^* Using the neighbourhoods we next define a class of functions ^ i , as follows:

where it is indexed by a row vector-valued function yn(x)ern, aeja^, which is also implicit in Ki,(), indicator function TjeJ. Let y^(Ay) = c, [M,,(A'y)]-', y,(.^O) = e, [ M \

and an

and

since Ri(xa)=gn{ei,X,,xo)-gM(e,,X,,xo), the result follows if two conditions are met: (1) equicontinuity of the process I J . , S^., ?(,, X,,Xj) over ^ , , in a neighbourhood of ^^(e,, ^ , , Xj) and (2) that, with probability approaching 1, (,,X,,Xj) lies within the neighbourhood over which we establish equicontinuity. We use the jSf^-norm to examine (1). We verify both of these two conditions in tum. We verify the equicontinuity condition (1) using a lemma in Ichimura (1995)." We first define some notation in order to state his lemma. For r = 1 and 2, let 3C' denote the r-fold product space of 3^<=/?'' and define a class of functions i^ defined over SC'. For any v^e*?,, write v,./, as a short hand for either v,(x,) or i^n(x,,, x,j), where ii ?ti2. We define {/ Vn = l]j V'n.*,. where X, denotes the summation over all permutations of r elements of { x i , . . . , X n } f o r r = l or 2. Then {/ Vn is called a U-process over (f/nST,. For r = 2 we assume that \^n(Xi,Xj)=\yn(Xj,X,). Note that a normalizing constant is included as a part of v'n- A U-process is 24. Note that a calculation using change of variables and the Lebesgue dominated convergence theorem shows that on 5, [A/,n(x)]~' converges to a nonsingular matrix which only depends of K(-) times [l//(.x)l'*'. Hence, on S, each element of [Mp,,(x)Y^ is uniformly bounded. Thus use of the sup-norm is justified. 25. The result extends Nolan and Pollard (1987), Pollard (1990), Arcones and Gine (1993), and Sherman (1994) by considering U-statistics of general order r g 1 under inid sampling, allowing ^ to depend on n. When 3^ depends on n, we need to assume condition (ii) in the lemma below as noted by Pollard (1990) when r=\.

286

REVIEW OF ECONOMIC STUDIES

called degenerate if all conditional expectations given other elements are zero. When r = l , this condition is defined so that E(l|/) = O. W e assume t h a t ',cSe\3>>'), where <e\0>') denotes the :Sf'-space defined over x' using the p r o d u c t measure of ^ , ^ \ W e denote the covering n u m b e r using i f ' - n o r m , || II2, by N2(e, ^, * ) . "

Lemma 3 (Equicontinuity). Let {X,)".^ be an iid sequence of random variables generated by 9. For a degenerate U-process {Ul|/} over a separable class offunctions ^'<^^\^') suppose the following assumption hold {let \\vA2 = [i:,^^{w.j,}]"y, (i) There exists an F^^<e\9') such that for any l|/e',\y/\<F such that lim sup,^

ZE{FlJ

(ii) For each e>0, lim,^ l^^ ^{Flj, I{F,.,,> e}} = 0; (iii) There exists X(e) and e>0 such that for each e>0 less than e.

and \l [log X (x)\'''^dx < 00. Then for any e>0, there exists S>0 such that limPr{
\ \

sup
h S S

|f4(v,,-V2)l >e} = 0.

Following the literature we call a function FJ^ an envelope function of .^n if for any l|/e^, y/nj ^F, holds. In order to apply the lemma to the process XJ., Z;.,^,(,,X, A}) over ^,, in a neighbourhood of go(e,, XI, XJ), wefirstsplit the process into two parts; the process Yjg,(i, X,, Xj) = '^ g(,ei, X,, EJ, XJ), where gl(e,, X,, sj, XJ) = \g,(e,,X,, Xj) +gAej,Xj,X,)]/!,

and the process i ; . , ^(,,X,,^,). Note that g(eX,,X,) = n-^'^ Yn(X,) e\ er (ah^)-''K{0) 7, is a order one process and has mean zero, hence it is a degenerate process. On the other hand g{ei,Xi, ej,Xj) is a order two process and not degenerate, although it has mean zero and is symmetric. Instead of studying g'!,(e,,Xi,ej,Xj) we study a sum of degenerate U-processes following Hoeffding (1961)." Write Z,= (e,. A-,), (^(Z,) = E{g(Z,, z)|Z,} = E {g(r, Z,) |Z,}, and gl(Z,, ZJ) =g'i,{Z,, ZJ) - MZ,) Then MZj).

where g(Zi, Zj) and 2 (/i- 1) (|>(Z|) are degenerate U-processes. Hence we study the three degenerate Uprocesses: |;(Z,,Z,),2- ( n - l ) - 0,(Z,), and g(e,,Xl,X|), by verifying the three conditions stated in the equicontinuity lemma. We start by verifying conditions (i) and (ii). An envelope function for g(Z,, Zj) can be constructed by the sum of envelope functions for g(e,,X|,X|) and g,(e,,A',,A'y). Similarly an envelope function for i"(Zi, ZJ) can be constructed by the sum of envelope functions for gJ(Z,, Zy) and 2 0n(Z,). Thus we only need to construct envelope functions that satisfy conditions (i) and (ii) for g^S|,X|,X|), gn{e,,X,,Xj) and 2n<|>(Zl). Let l7 = \{fx(X,)^qa-2Ef} for some 9o>2/>0. Since sups|/(;c)-/fWI ge/, If-^l holds for any 7,6J*. Also for any neighbourhood of [A/^(x)] ' defined by the sup-norm, there exists a C>0 such that \[M,{x)Y'\^C so that \g{eX,,X,)\^n-"^C \s,\ l{ao-S)h^''K(0) If and the second moment of the right hand side times n is uniformly bounded over n since the second moment of e, is finite and nht-<-oa. Hence condition (i) holds. Condition (ii) holds by an application of Lebesgue dominated convergence theorem since nh^-* 00. 26. For each e>0, the covering number N,(e,^,^) is the smallest value of m for which there exist functions ?,,...,?, (not necessarily in ^) such that min, [E{|/-g,|'} "'\ g e for each / in S^. If such m does not exist then set the number to be 00. When the sup-norm is used to measure the distance in calculating the covering number, we write N^{e, ^). 27. See also Serfling (1980).

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

287

Note that any element of [[{X,-XJ)/(ah)\'\'KH{X,-X)) l{\\X,-Xj\\<.C2-h^} for some C, and Cj.Thus ^-\ e,\ CQ-

is bounded by C^[{ao-S)h\-' if.

Hao-Sa)h]-'I{\\X,-XJ\\^C2h}

Therefore, analogous to the previous derivation, conditions (i) and (ii) hold for gn{Si,Xi,Xj). Note further that since the density of x is bounded on 5 by some constant, say, Ci>Q,\4>n(et,X,)\i. n'^''^-\e,\ C- Ci- Cj and hence 2 n | (^(e,,X)| has an envelope function w''^^2 | e,| C- C^- Cj that satisfies the two conditions. To verify condition (iii), first note the following. Write Ji,(X,-Xj)-^[[(X,-Xj)/{aK)Y'];KH{X,-Xj) and Xj) = [[(X-Xj)/(aoK)f'^K^(Xi-Xj). Using this notation ,| \r^(Xj) MX-Xj) Ij If (L-3) If

E>\ \y(XJ)-Y.o(XJ)\ C, l(ao-Sa)h]-''I{\\X-XJ\\SC2h} ^\ e,| \rno(Xj)\ \MX!-Xj)-JUX,-Xj)\ Xj)\ -iTj-Ijl. For g{e,,X|,X|) the right hand side is bounded by some O O ,

Since nA^->oo, the if2-covering number for the class of functions denoted by g,(ei,Xi,X,) for^.eS?], can be bounded above by the product of the covering numbers of F,, si, and J. Since it is the log of the covering number that needs to be integrable, if each of three spaces satisfy condition (iii), then this class will also. Cleariy .s^ satisfies condition (iii). To see that F, and J do also, we use the following result by Kolmogorov and Tihomirov (1961)." First they define a class of functions for which the upper bound of the covering number is obtained. Definition 2. A function in'ViK) has smoothness q>Q, where q=p + a with integer p and 0 < a g 1, if for any xeK and x + heK, we have
/!, X),

where Bt(h, x) is a homogenous form of degree kinh and \R^(.h, x)\^C\\h\\'', where C is a constant.

LetIf(C)

{v()|^(,)|g||r}

If a function defined on K is p-times continuously differentiable and the p-th derivative satisfies Holder continuity with the exponent 0< a g 1, then a Taylor expansion shows that the function belongs to f* (C) for some C, where q=p + a. Lemma 4 (K-T). For every set Ac:^'^(C), where KczR'', wehave,forO<d,q<co, Iog2 N^(e, A)^L(d, q, C, K)(\/e)'"'' for some constant L(d, q,C,K)>0. Hence, because d/q<\, condition (iii) holds for F, and J. Analogously we can verify condition (iii) for the remaining U-processes. Hence all three processes are equicontinuous. The remaining task is to verify that {,,Xi,Xj) lies in the neighbourhood of gM(ei,Xi,Xj) over which we showed equicontinuity. By the inequality in Lemma 3 (L-3), this follows from Assumptions 3 and 5, and by verifying that almost surely sup
xeS.aej^

\\M,(x)-M,(x)\\->q,

where limn^ inf^^sdet (M,,(x))>0." The latter follows directly from the nonsingularity of matrix Af, and the trimming rule. Hence the following lemma completes the proof. 28. See pp. 308-314. Kolmogorov and Tihomirov present their result using the concept of packing number instead of covering number. 29. Recall from the discussion of Theorem 3 that M,(x) depends on the parameter a.

288
Lemma 5.

REVIEW OF ECONOMIC STUDIES


Under the assumptions of Theorem 3, almost surely, sup
xeS.aej/

\\M,^x)-M,(x)\\-^O.

Proof

Note that any element of the matrix difference Mp (x) - Mp (x) has the form

where in the notation introduced in Section A.I the kernel function G(s) = ^ ^K(s) for some vectors q and r whose elements are integers and they sum to p or less, where q and r depend on the element of M being examined. To construct a proof, we use the following lemma of Pollard (1984).^" Lemma 6 (Pollard). For each n, tet ^ be a separable class offunctions whose covering numbers satisfy supN,(e,&>,'i>)^Ae''*' for 0<e<\

with constants A and W not depending on n. Let {^} be a non-increasing sequence of positive numbers for which lim,^ nC,,l/\ogn = co. / / | vl g 1 and (E{v'} f'^'^l^^for each y/ in ", then, almost surely.

To use this lemma, we need to calculate N,(e, 9, "V,). Let C, Ssup,e G(s), Cj is a Lipschitz constant for G, and Cj is a number greater than the radius of a set that includes the support of G. In our application, recall from our proof of Theorem 3 that .a^ = [oo - 5 , , oo + 5.1, where 0 < S. < Oo so that

The upper bound of the right hand side does not depend on 3. Moreover, the right hand side can be made less than E C for some C> 0 by choosing I a i - Oj | g e and | xo, - Xo21 g e. Since S and .aT are both bounded subsets of a finite dimensional Euclidean space, the uniform covering number condition holds. To complete the proof of Lemma 5, note that we are free to choose ^, = 1 and f = C Af ^ in our application of the lemma. || Next we examine the second term (B-3). Lemma 7 (Term (B-3)). where Under the assumptions of Theorem 3

"~'^^Z"_i R2(X,) = 0p(\), and Ri^Xo) is defined as the difference between Term (B-3) and b{xo). Proof. Note that (B-3) = c, lM,Axo)V'n-'H'X',ixo)Wixo)J^f(xo)^*{xo) I = e, [M,(x)]- I f . , , , - i ; . , ll(X,-xo)/(ah)]<^']'^X,-xo)'W^\xo)K,(X,-xo) / = e, [M,(xo)]" I f . , , , -' i ; . , mX,-Xo)/(ah)]']'iX,-Xo)'''^'K,(X,-Xo) (L-7A) /. (L-7B)

- E {l[(X,-Xo)/(ah)f']'(X^-Xo)'^">K,{X,-Xo)\xo}W'\xo) I + e, [M^Axo)]-' I f . , , , mliX,-xo)/(ah)f']'{X,-Xo)"'K,iX,-Xo)\x}m'^\xo)

30. His lemma only requires that .^n is a permissible class of functions. In our applications ^ is always separable. His Example 38, pp. 35-36, gives a similar result. We provide a proof here for completeness and for later reference.

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

289

Define term (L-7A) as Ri\{xo). We apply the same method as in Lemma 2 to show that n'"^ Z"_i ^21 (X) = Op(l). Instead of 'S^, define the class of functions ^ i , as follows:

where it is indexed by a row vector-valued function f(x)eT, ae.s^, which is also implicit in Ki,(), and an indicator function ijSJ. Let
j X,-XJ) Ij,

and To prove that term (L-7B) equals b(xo) + o{hi) we use the assumption that all the moments of Ar(-) of order p +1 and higher up to p have mean zero, that Xo is an interior point of the support of x that is more than (ao + 5 )/! C interior to the closest edge of the support, where C is the radius of the support of Ar( ), and the assumption that the density of x is ^-times continuously differentiable and its p-th derivative satisfies a Holder condition. Using a change of variable calculation and the Lebesgue dominated convergence theorem, and Lemma 5 the result follows. || Note that this term converges to zero with the specified rate only if xo is an interior point. Thus for kernel regression estimators for higher dimensions, we need to introduce a special trimming method to guarantee this. Use of higher order local polynomial regression alleviates the problem for higher dimensional problems, but at the price of requiring more data locally.
Lemma 8 (Term (C-3)). Under the assumptions of Theorem 3,

(C-3) = o,(/.j). Proof. Recall that the third term equals e, [Mpn(xo)]'''n''H'X'p(xo)lVixo)rp(xo) h. Note that W(xo)rf{xo)Io\\ "-1 [(.X-Xo)/(ah)]^'1{X,-xo)/hf'^\m'f\x,) -m'^\xo)]K,(xo-X>)/ij ||

where the inequality follows from the Holder condition on m'^'(xo) and the compact support condition on K( ), and the last equality follows from the same reasoning used to prove Lemma 5. The conclusion follows from Lemma 5 and the assumption of nonsingularity of Mp(xo). || Lemmas 2-8 verify Theorem 3. Corollary 1 in the text follows as a special case for /? = 0. A.3. Verifying the assumptions of Lemma 1

Five assumptions in Lemma 1 are as follows: (i) Both P(z) and g(t, p) are asymptotically linear with trimming where

lg{t,p)-g(t,p)]I(xeS)

= n-' i ; . , y/,^YJ, Tj, P{Zj);

(ii) dg(t,p)/dp and P{z) converge uniformly to dg{t,p)/dp and P{z), respectively, and that dg(l,p)/dp is continuous for all ( and p; (iii) p\im^^n-"^Z';.,b,(.Tl,XJ) = b, and plim_-'^='S;., dg(T,, P(Z,))/8p b,{T,, P(Z<)) = b,^\ (iv) plim^-/^S;., [di(T,,Pr,(Z,))/Sp-dg(T,,P(Z,))/dp]-R,(Z,) = (i(V) plim_-'^^E;-, E ; . , [dg(T,,Pr,(Z,))/8p-dg(T,,P(Z,))/8p] v',(D;, Z^; Z,) = 0. We verify these assumptions for the local polynomial regression estimator. Condition (i) is just a consequence of Theorem 3. We next show that the derivative of the local polynomial regression estimator converges uniformly to the derivative of the limit of the local polynomial regression estimator.

290

REVIEW OF ECONOMIC STUDIES


Theorem 4. If Assumptions 1-4, and 8 hold, then dg(t,p)/dp is uniformly consistent for dg{t,p)/dp.

Proof For convenience we drop the subscripts p, p, and the argument Xo of X,, (xo), Pp (xo), P* (xo), and ^(xo) here so thatX=Xp(xo), P = Pp(xo), P* = P*{xo), and_ W= ^(xo). Also denote the derivative with respect to the first argument of xo by V. Note that X'fVY=X'lVXp. Hence by the chain rule.

Since

{^(X-W)-lV(X'fVX)](X']VX)-'X'W}xp*

we obtain

Note that for.sgl.

= [((A:-XO)<" "')+..-..,-,, 0 , . . . , 0], where the second equality follows from our convention on the order of the elements. Thus

VX=-{0

1 0 . . . 0 (;t-xo)e"> 0 . . . 0 . . . ( x - x o ) ^ " - " 0 . . . 0).

Note that each column of VX is either a column of X or the column with all elements being 0. Hence there exists a matrix J such that VX=-XJ, where 7 is a matrix that selects appropriate column of A" or the zero column. Without being more specific about the exact form of J we can see that

-(X'fVXy'(X'fVVX)P* = Jp*, and also that

Next, to simplify the last expression, we use some specific properties of matrix J. The key properties of J we use are that all elements of the first column are 0 and that thefirstelement of the second column of 7 is I. That is, thefirstcolumn of AJ is always the 0-vector, regardless of A and the second column of AJ is the first column of A. Since the first column is chosen by J exactly once, the preceding observations also imply that thefirstrow of 7 is ej, where e2 = (0, 1,0,. . . , 0) if p^ 1 and 0, if p = 0. Therefore

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

291

where ei JP* = Vm(x). That the remaining two terms converge uniformly to zero can be shown analogously as in Lemma S. || Condition (iii) of Lemma 1 clearly holds under an i.i.d. assumption given the bias function defined in Theorem 3. In order to verify condition (iv) of the lemma recall the definition of the residual terms and use the same equicontinuity argument as in the proof of Theorem 3. Finally condition (v) can be verified by invoking the equicontinuity lemma. This is where the additional smoothness condition is required. Armed with these results, wefinallyturn to the proof of key Theorem 2. A.3. Proof of Theorem 2 Note first that, writing I,=I(X,eS),

We first consider the numerator and then tum to the denominator of the expression. Note that the numerator of the right-hand side of (T-1) is the sum of three terms (TR-l)-(TR-3): writing g, (x) = E5(y,|Z)= LA"=x).

Terms (TR-1) and (TR-2) are analogous to terms we examined in Theorem 1. Term (TR-3) is the additional term that arises from estimating g{x). However, the first two terms from Theorem 1 have to be modified to allow for the trimming function introduced to control the impact of the estimation error of g(x). Central limit theorems do not apply directly to the sums in (TR-1) and (TR-2) because the trimming function depends on all data and this creates correlation across all i. Instead, writing I, = I(XieS), we show that these terms can be written as

and

respectively. One can use the equicontinuity lemma and our assumption of p-nice trimming to show the result for term (TR-1). The same method does not apply for term (TR-2), however. This is because when we take an indicator function 7/ from J, where /,#/,, then

It is necessary to recenter this expression to adjust for the bias that arises from using 7,. In order to achieve this we observe that, writing As(A',) = [gi(A'i)-g(A',)]-Es(yi- yol-D=l),

where d(A',) = |/(A',)-/(A',)|, ^-(.s) = 1 if-1 g.s<0 and 0 otherwise, and ^+(.s)= 1 if Ogj< 1 and 0 otherwise. Since/(X,) = (iVoa)"''E^^, KiXXi-Xi)lan), the latter two terms can be expressed as double sums. We then apply an equicontinuity argument to the expressions

292
and

REVIEW OF ECONOMIC STUDIES

and control the bias by ^-smoothness of/(A",). Finally, term (TR-3) can be written as the sum of three terms ^,^ .^,^ V<^MYoj,Xj;X,), (TR-3-1) (TR-3-2) and (TR-3-3) Terms (TR-3-2) and (TR3-3) are 0,(1) by the definition of asymptotic linearity of ^(;ir,). Term (TR-3-1) is a U-statistic and a central limit theorem can be obtained for it using the lemmas of Hoeffding (1948) and Powell, Stock, and Stoker (1989) and a two-sample extension of the projection lemma as in Serfling (1980). We first present the Hoeffding, Powell, Stock, and Stoker result. Lemma 9 (H-P^S). Suppose {Z,}?., is i.i.d., Uy/=in ( n - 1))-' J,. V(Z,, Z,), where ^,{Z,, Zj) = Vn{Zj,Z,) and V.{v^(Z,,Zj)}=Q, and & V-. = ' ' " ' I ^ . . , 2 ;7,(Z,), H-Zie/-/ ;,,(Z,) = E{v,,(Z,, Z,)|Z,}. / / ^{^l^(Zi,ZJf}=o(n), then /IE[(t/ v , - [ / ,/')'] = o(l). We also make use of the results of Serfling (1980). Lemma 10 (Serfling). Suppose {Zaj},,,^ and {Z,,,},^,, are independent and within each group they are i.i.d., Un,.., Vn,.n, = ("o ' ,)^' !, I,,,, w^.., (Zoi, Z,j), and E{ ^ ^ , , (Zoi, Z,j)} = O,and t4o., W.o.m = no ' I,,,,/'o^o.,, (Zo,) -I- nr' I,.,,, p,^.^, (Z,j), where for k = Q,\, p,^,^ (Z,j) = E{ ^>^,^^ (Zo,, Z,,) |Z*,}. // 0<lim^ ,/=;;< 00, where = + /,, and E{v^^,,(Zo,, Z,^)^} =o(no) + o(n,), then E[(f/ U )'] (l) In order to apply the lemmas to term (TR-3-1), note that it can be written as ^ r ' ^ ' I,^,, I,.^,, j ^ , V i A T j A T , ( Y,j, XJ ; X,) + A^r'^'Z,e,, Vi.Ar,(r,,,X,;.y,) + N^"^ No' E,,,, Y.J,,, VowoAT,(Ko,,XJ;.V,). Term (TR-3-la) can be rewritten as Afr'^' I , , , , E,.^,, ^^. v?;vo, ( Y,j, Xj; ^ , ; Y,,, X,; A^^) where Wif^o^, ( J'ly, A-y; X,; X,,, A-,; A^,) = [y,,^^,,, (K,,, A^,; X ) + VI^VOA-, (X,,, A",; Ay)]/2. Thus by Lemma 9 and assumption (ii-a), term (TR-3-la) is asymptotically equivalent to A^r'-" I , , , , E {v/,;v<,/v, ( Yu,X< ; XJ)I y,,, AT,}. By assumption (ii-a), term (TR-3-Ib) is o,(l). By Lemma 10 and assumption (ii-a), term (TR-3-lc) is asymptotically equivalent to
j ^ A-, ( YOJ,
XJ

(TR-3-la) (TR-3-lb) (TR-3-lc)

; X,)\ YOJ, XJ}.

Hence putting these three results together, term (TR-3-1) is asymptotically equivalent to

Collecting all these results, we have established the asymptotic normality of the numerator. We next examine the denominator of (T-1). Note that N^' E,^,^ ii=N'[' E_^, I, + NT' .^, ( / , - / , ) . The first term on the right-hand side converges in probability to E(/) by the law of large numbers. To see that the second term is o,(l), note that Afr'll,,,,(7,-/,)| g A T ' I , , , , | / , - / , | . The Markov inequality implies for any

HECKMAN ET AL.

MATCHING AS AN ESTIMATOR

293

e>0, Pr {AT' X,^, |7i-/,|>}gE{|/,-/,!}/. Hence assumption (ii-d) implies that the second term iso,(l). This result, in conjunction with our result for the denominator, proves Theorem 2. ||

Acknowledgements. The work reported here is a distant outgrowth of numerous conversations with Ricardo Barros, Bo Honore, and Richard Robb. We thank Manuel Arellano and three referees for helpful comments. An eariier version of this paper "Matching As An Evaluation Estimator: Theory and Evidence on Its Performance applied to the JTPA Program, Part L Theory and Methods" was presented at the Review of Economic Studies conference on evaluation research in Madrid in September, 1993. This paper was also presented by Heckman in his Harris Lectures at Harvard, November, 1995, at the Latin American Econometric Society meeting in Caracas, Venezuela, August, 1994; at the Rand Corporation, U.C. Irvine, U.S.C., U.C. Riverside, and U.C. San Diego September, 1994; at Princeton, October, 1994; at UCL London, November, 1994 and November, 1996, and at Texas, Austin, March, 1996 and at the Econometric Society meetings in San Francisco, January, 1996. REFERENCES ARCONES, M. A. and G I N 6 , E. (1993), "Limit Theorems for U-processes", Annals of Probability, 21, 14941542. BARNOW, B., CAIN, G. and GOLDBERGER, A. (1980), "Issues in the Analysis of Selectivity Bias," m E. Stromsdorfer and G. Farkas (eds.). Evaluation Studies Review Annual, Volume 5 (San Francisco: Sage). BARROS, R. (1986), "Nonparametric Estimation of Causal Effects in Observational Studies" (University of Chicago, mimeo). COCHRANE, W. G. and RUBIN, D. B. (1973), "Controlling Bias In Observational Studies", Sankyha, 35, 417-446. DAWID, A. P. (1979), "Conditional Independence in Statistical Theory", Journal of The Royal Statistical Society Series B, 41, 1-31. FAN, J. (1993), "Local Linear Regression Smoothers and Their Minimax Efficiencies", The Annals of Statistics, 21, 196-216. FAN, J., HARDLE, W. and MAMMEN, E. (1996), "Direct Estimation of Low Dimensional Components m Additive Models" (Working paper). HAHN, J. (1996), "On the Role of the Propensity Score in the Efficient Semiparametric Estimation of Average Treatment Effects" (unpublished manuscript. University of Pennsylvania). HECKMAN, J. (1974), "Shadow Prices, Market Wages, and Labor Supply", Econometrica, 42, 679-694. HECKMAN, J. (1990), "Varieties of Selection Bias", American Economic Review, 80, 313-318. HECKMAN, J. (1997), "Instrumental Variables: A Study of the Implicit Assumptions Underlying One Widely Used Estimator for Program Evaluations", Journal of Human Resources, 32, 441-462. HECKMAN, J., ICHIMURA, H., SMITH, J. and TODD, P. (1996a), "Nonparametric Characterization of Selection Bias Using Experimental Data, Part II: Econometric Theory and Monte Cario Evidence" (Unpublished manuscript. University of Chicago). HECKMAN, J., ICHIMURA, H., SMITH, J. and TODD, P. (19966), "Sources of Selection Bias in Evaluating Programs: An Interpretation of Conventional Measures and Evidence on the Effectiveness of Matching As A Program Evaluation Method", Proceedings of The National Academy of Sciences, 93, 13416-13420. HECKMAN, J., ICHIMURA, H., SMITH, J. and TODD, P. (1998), "Characterizing Selection Bias Using Experimental Data", Econometrica (forthcoming). HECKMAN, J., ICHIMURA, H. and TODD, P. (1997), "Matching As An Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Program", Review of Economic Studies, 64, 605-654. HECKMAN, J. and ROBB, R. (1985), "Altemative Methods For Evaluating The Impact of Interventions", in J. Heckman and B. Singer (eds.). Longitudinal Analysis of Labor Market Data (Cambridge: Cambridge University Press). HECKMAN, J. and ROBB, R. (1986), "Altemative Method For Solving the Problem of Selection Bias in Evaluating The Impact of Treatments on Outcomes," in H. Wainer (ed.). Drawing Inferences from SelfSelected Samples (New York: Springer-Veriag). HECKMAN, J. and SMITH, J. (1998), "Evaluating The Welfare State", Frisch Centenary, Econometric Monograph Series (Cambridge: Cambridge University Press), forthcoming. HECKMAN, J., SMITH, J. and CLEMENTS, N. (1997), "Making the Most Out of Social Experiments: Reducing the Intrinsic Uncertainty in Evidence From Randomized Trials With An Application to the National JTPA Experiment", Review of Economic Studies, 64, 487-535. ICHIMURA, H. (1993), "Semiparametric Least Squares (SLS) and Weighted SLS Estimation of Single Index Models", Journal of Econometrics, 58, 71-120. ICHIMURA, H. (1995), "Asymptotic Distribution of Nonparametric and Semiparametric Estimators with Data-Dependent Smoothing Parameters" (Unpublished manuscript. University of Chicago). KOLMOGOROV, A. N. and TIHOMIROV, V. M. (1961), "e-Entropy and -Capacity of Sets in Functional Spaces", American Mathematical Society Translations, series 2, 17, 277-364.

294

REVIEW OF ECONOMIC STUDIES

LINTON, O. and NIELSON, J. P. (1995), "A Kernel Method of Estimating Structural Nonparametric Regression Based on Marginal Integration", Biometrika, 82, 93-100. NOLAN, D. and POLLARD, D. (1987), "U-processes: Rates of Convergence", Annal of Statistics, 15, 405414. NEWEY, W. K. and McFADDEN, D. L. (1994), "Large Sample Estimation And Hypothesis Testing", in R. F. Engle and D. L. McFadden (eds.). Handbook of Econometrics. Vol. IV. (Amsterdam: Elsevier). MASRY, E. (1995), "Multivariate Local Polynomial Regression for Time Series" (Unpublished manuscript). POLLARD, D. (1990), Empirical Processes: Theory and Applications (Hayward: IMS). POWELL, J. L., STOCK, J. H. and STOKER, I. M. (1989), "Semiparametric Estimation of Index Coefficients", Econometrica, 57, 1403-1430. ROSENBAUM, P. and RUBIN, D. B. (1983), "The Central Role of the Propensity Score in Observational Studies for Causal Effects", Biometrika, 70, 41-55. RUPPERT, D. and WAND, M. P. (1994), "Multivariate Locally Weighted Least Squares Regression" The Annals of Statistics, 22, 1346-1370. SERFLING, R. J. (1980), Approximation Theorems of Mathematical Statistics (New York: Wiley). SHERMAN, R. P. (1994), "Maximal Inequalities for Degenerate U-processes with Applications to Optimization Estimators", Annal of Statistics, 11, 439-459. STONE, C. (1982), "Optimal Global Rates of Convergence For Nonparametric Regression", Annals of Statistics 10, 1040-1053. TJOSTHEIM, D. and AUESTAD, B. H. (1994), "Nonparametric Identification of Nonlinear Time Series: Projections", Journal of the American Statistical Association, 89, 1398-1409. WESTAT, INC. (1980), "Net Impact Report No. 1, Impact on 1977 Eamings of New FY 1976 CETA Enrollees in Selected Program Activities" (Rockville, Maryland, Westat Inc.). WESTAT, INC. (1982), "CLMS Follow up Report No. 7 (18 Months After Entry), Postprogram Experiences and Pre/Post Comparisons For Terminees Who Entered CETA During July 1976 Through September 1977" (Rockville, Maryland, Westat Inc.). WESTAT, INC. (1984), "Summary of Net Impact Results" (Rockville, Maryland, Westat Inc.).